Signal, noise, and bias in phylogenetic inference:potential and limits to the resolution of phylogenetic trees in the phylogenomic era

This paper establishes a theoretical framework demonstrating that while phylogenetic signal and bias accumulate linearly with data size, stochastic noise accumulates nonlinearly, revealing that genome-scale datasets may still fail to resolve deep or rapid divergences when noise overwhelms signal or when systematic bias continuously outpaces signal accumulation.

The Gist

The Big Idea: Why More Data Doesn't Always Mean a Better Answer

Imagine you are trying to solve a massive jigsaw puzzle to figure out the family history of all living things (the "Tree of Life"). In the past, scientists only had a few puzzle pieces. Now, thanks to modern technology, we have millions of pieces.

You might think, "If I have a million pieces, I can definitely solve the puzzle!" But this paper argues that having more pieces doesn't guarantee a clear picture. Sometimes, adding more pieces just makes the picture more confusing.

The authors break down the puzzle pieces into three distinct categories: Signal, Noise, and Bias.

---

1. The Three Players in the Game

Signal: The True Clues

• What it is: These are the puzzle pieces that actually tell the truth about who is related to whom. They are the "right answers."
• How it grows: Signal grows linearly. Imagine a straight line going up. Every time you add a good piece, you get a little bit closer to the solution. It's steady and reliable.

Noise: The Static on the Radio

• What it is: This is random confusion. It's like static on a radio or random scratches on a puzzle piece that happen to look like they fit but don't. It's caused by pure chance (random mutations).
• How it grows: Noise grows non-linearly (it curves). At first, when you have very few pieces, the noise is huge compared to the signal. It's like trying to hear a whisper in a hurricane.
• The Good News: As you keep adding pieces, the "randomness" starts to cancel itself out. The curve flattens. Eventually, if you have enough pieces, the steady signal should overtake the random noise.
• The Catch: Sometimes the "signal" is so weak (like a very faint whisper) that even with a million pieces, the noise never fully goes away. You might never solve that specific part of the puzzle.

Bias: The Tricky Forger

• What it is: This is the most dangerous player. Bias isn't random; it's systematic. Imagine a forger who deliberately paints fake clues on the puzzle pieces to trick you into thinking two unrelated people are cousins.
• How it grows: Bias also grows linearly, just like Signal.
• The Danger: If the "Forger" (Bias) is working harder than the "Truth-teller" (Signal), adding more pieces just gives the forger more chances to lie. You can add a billion pieces, but if they are all biased, you will confidently build the wrong tree.

---

2. The "More Data" Myth

For a long time, scientists believed in a simple rule: "If the answer isn't clear, just get more data." They thought, "Signal will eventually beat Noise."

This paper says: "Not always."

• Scenario A (The Solvable Puzzle): You have a deep family split (like humans vs. fish). The signal is strong. Even if there is noise, adding more data eventually drowns out the noise, and you get the right answer.
• Scenario B (The Impossible Puzzle): You have a very recent split (like two cousins who look identical). The "signal" is incredibly faint. The "noise" is loud. Even if you scan the entire genome, the signal might never be strong enough to overcome the noise. You are stuck.
• Scenario C (The Trap): You have a "Forger" (Bias). Maybe two unrelated animals both evolved to have a lot of the same DNA letters just by chance (like both evolving to be very fast). The forger is lying so convincingly that no matter how many pieces you add, the picture stays wrong.

---

3. Real-World Examples from the Paper

The authors tested this theory on two real-world datasets:

1. Birds (The Hoatzin): Scientists have argued for years about where a weird bird called the Hoatzin fits in the bird family tree.
   • The Result: The authors found that for this specific bird, almost every piece of DNA they looked at had more noise than signal. It wasn't that the data was "biased" (lying); it was just too "noisy" (confusing). The puzzle pieces were too blurry to see the picture.
2. Fish (Sleepers): They looked at a group of fish where the family tree was also confusing.
   • The Result: They found that many of the DNA markers (called UCEs) that scientists usually trust were actually full of noise. In fact, if they added the "noisiest" pieces first, they would need 110,000 characters just to start seeing the truth. If they had picked the "cleanest" pieces first, they would have solved it much faster.

---

4. The Takeaway: Quality Over Quantity

The main lesson of this paper is that not all data is created equal.

• Don't just dump more data: Throwing a million random puzzle pieces at a problem won't help if half of them are blurry (noise) or fake (bias).
• Be a detective: Before you start a study, you need to predict which pieces will give you the "Signal" and which will just give you "Noise" or "Bias."
• The Future: We need to design experiments that specifically hunt for the "Signal" and avoid the "Forgers." We need to be smart about which genes we sequence, not just how many.

In short: In the age of big data, we can't just rely on volume to solve mysteries. We have to understand the nature of the data. Sometimes, the answer isn't hidden because we don't have enough information; it's hidden because the information we have is too noisy or too tricky to trust.

Technical Summary

1. Problem Statement

Despite the advent of phylogenomics, where datasets routinely span thousands of loci and millions of characters, many deep branches in the Tree of Life remain unresolved or exhibit strongly supported but conflicting topologies. The central problem addressed is the assumption that "more data equals better resolution." The authors argue that not all data are equally informative; some loci may be deceptive due to the complex interplay of three distinct components:

• Phylogenetic Signal: True synapomorphies reflecting common ancestry.
• Stochastic Noise: Random homoplasy (parallelism/convergence) due to chance.
• Systematic Bias: Non-random, lineage-specific convergence (e.g., due to compositional heterogeneity or biased gene conversion) that systematically misleads inference.

Existing metrics (e.g., site-rate estimates, saturation indices) are retrospective and lack a predictive theory for how these components scale and interact as character sampling increases.

2. Methodology

The authors derive a general analytical framework based on previous theoretical work (Townsend et al., Su et al.) to model the expected accumulation of signal, noise, and bias as a function of character sampling ($n$).

• Theoretical Model:
  • Signal ($Y$): Modeled as the cumulative number of characters supporting the correct quartet topology via true synapomorphy. The authors demonstrate that signal accumulates linearly with character sampling ($E[Y] \propto n$).
  • Noise ($W^*$): Modeled as the net support for incorrect topologies arising from stochastic homoplasy. The authors derive that noise accumulates nonlinearly with a concave trajectory. This is driven by a square-root term ($\sqrt{n}$) arising from the "random walk" competition between the three possible quartet resolutions.
  • Bias ($B$): Defined as the difference between total noise with compositional heterogeneity ($W$) and noise without it ($W^*$). Bias is shown to accumulate linearly ($E[B] \propto n$).
• Empirical Validation:
  The theory was applied to two large-scale empirical datasets:
  1. Avian Phylogeny: 259 loci from Anchored Hybrid Enrichment (AHE) focusing on the placement of the Hoatzin (Opisthocomus hoazin).
  2. Acanthomorph Fish Phylogeny: >1,001 Ultraconserved Element (UCE) loci focusing on the placement of sleepers (Kurtidae).
• Simulation & Analysis: The authors calculated expected signal, noise, and bias for individual loci and cumulative datasets, testing different sorting strategies (e.g., adding loci by highest signal-to-noise ratio vs. worst noise-to-signal ratio).

3. Key Contributions & Theoretical Insights

A. Divergent Scaling Properties

The paper establishes a fundamental mathematical distinction in how phylogenetic components scale:

• Signal: Linear accumulation ($y = mx$).
• Noise: Nonlinear, concave accumulation ($y \approx \sqrt{x}$ initially, becoming linear asymptotically but with a lower slope than signal in ideal cases).
• Bias: Linear accumulation ($y = mx$).

B. The Limits of "Sampling One's Way Out"

The authors challenge the conventional wisdom that increasing data volume will inevitably resolve phylogenetic conflicts:

• Signal vs. Noise: While signal generally grows faster than noise, if the internode (branch length between speciation events) is extremely short or deep, the slope of signal accumulation can be so shallow that even genome-scale data may never practically exceed the noise floor.
• The Danger of Bias: Because bias accumulates linearly, it can have a slope steeper than that of the signal. In such cases, adding more data will never resolve the conflict; instead, it will continuously reinforce the incorrect topology. This renders the "more data" strategy futile for problems dominated by systematic bias.

C. Character-Acquisition Bias vs. Phylogenetic Bias

The paper distinguishes between:

• Character-Acquisition Bias: Uniform constraints (e.g., codon usage bias, elevated AT content across all lineages) that reduce the effective state space dimensionality. This amplifies stochastic noise but does not necessarily create directional phylogenetic bias.
• Phylogenetic Bias: Heterogeneous biases among lineages (e.g., different lineages converging on the same base composition) that create systematic, directional errors.

4. Key Results from Empirical Datasets

• Avian AHE Dataset (Hoatzin):

  • For the internode determining the Hoatzin's placement, every single locus contributed more noise than signal.
  • Phylogenetic bias was negligible (<5% of loci showed bias > signal) because AT content was uniformly elevated across lineages.
  • Conclusion: The incongruence here is driven by noise, not bias. Resolution would require tens of thousands of characters to overcome the noise floor, regardless of locus selection order.

• Acanthomorph UCE Dataset (Sleepers/Kurtidae):

  • Across 1,001 loci, noise exceeded signal for the vast majority.
  • Locus Ordering Matters: The trajectory of resolution is highly sensitive to the order in which loci are added.
    • Adding loci sorted by Signal:Noise (best first) showed a convex trajectory, allowing signal to overtake noise relatively quickly.
    • Adding loci sorted by Noise:Signal (worst first) showed a concave trajectory, delaying the point where signal exceeds noise by requiring ~110,000 characters (a massive increase in data requirement).
  • Conclusion: A non-trivial fraction of loci actively impedes inference. Blind concatenation of all available data can be counterproductive.

5. Significance and Implications

• Theoretical Foundation: The paper provides the first predictive theory for the limits of accumulative character sampling. It reconciles the debate between "big data solves everything" and "some nodes are intractable" by showing that solvability depends on the relative slopes of signal, noise, and bias accumulation.
• Experimental Design: The framework shifts phylogenetic study design from intuition-based data collection to information-trajectory-based planning. Researchers can now predict whether a specific phylogenetic problem is solvable with current data types or if the internode is too short/biased to be resolved by sampling alone.
• Data Selection Strategy: The results argue against the indiscriminate concatenation of all available loci. Strategic selection (prioritizing high signal-to-noise loci) is critical. In some cases, the "best" strategy is to exclude specific loci that contribute more noise than signal.
• Bias Awareness: The work highlights that systematic bias is a "diabolical" hazard because, unlike noise, it cannot be overcome by simply adding more data if the bias slope exceeds the signal slope.

In summary, this paper moves phylogenetics beyond the "data deluge" mindset, offering a rigorous mathematical framework to diagnose why certain nodes remain unresolved and guiding researchers toward efficient, targeted data collection strategies that account for the distinct accumulation dynamics of signal, noise, and bias.